Repository for SciML AD backend types
Author SciML
11 Stars
Updated Last
7 Months Ago
Started In
February 2023

ADTypes.jl: Multi-Valued Logic for Automatic Differentiation Choices

Build Status

ADTypes.jl is a common system for implementing multi-valued logic for choosing which automatic differentiation library to use.

Why Should Packages Adopt This?

The current standard is to have a keyword argument with autodiff = true or autodiff = false. However, boolean logic is not sufficient to be able to do all of the choices that are required. For example for forward-mode AD there is ForwardDiff vs Enzyme, and thus true is ambgiuous. As libraries begin to support more and more of these autodiff mechanisms for their trade-offs, every library will have their own version of ChooseForwardDiff() etc. which would be a mess. Hence ADTypes.jl giving a single set of shared types for this. ADTypes.AutoForwardDiff() is the way to do it.

Why are they types instead of enums? Or EnumX?

If they were enums then they would not be dispatch type-level information. This is needed for the internals of many of the packages to be type-stable because the choice of config type is different per package, i.e. what needs to be cached in order to use everything in a non-allocating manner.

My AD Package is Missing?

Sure, give a PR to add it.

Required Packages

No packages found.